MTF-industrial-process

Data and R code to reproduce the results described in the paper "An Adaptive Learning Approach to Multivariate Time Forecasting in Industrial Processes" (Miguelez et al., 2024).

Data

This folder contains the dataset oee_data_4weeks.Rsd, a subset of the complete dataset comprising cleaned data from week 21 to week 24 of the year 2020, with the following variables:

• lsp: observation period id
• day: starting date
• hour: starting hour
• sh.id: work shift id
• new.sh: 1=first observation of the work shift, 0=otherwise
• wday: weekday
• tday: shift (Morning, Afternoon, Night)
• week.id: week id
• inistop: 1=the machine is inactive at the beginning of the period, 0=the machine is active
• of.id: production order id
• new.of: 1=first observation of the production order, 0=otherwise
• ref: reference
• ics: ideal cycle speed, units per minute
• TU: total units
• DU: defective units
• TgU: target units, OpT$\times$ics
• OT: period length, minutes
• SBT: stand-by time, minutes
• LT: loading time = OT-SBT
• rcs: real cycle speed = TU/LT
• lo: loading rate = LT/OT
• DT: downtime, minutes
• OpT: operating time = LT-DT
• av: availability rate = OpT/LT
• PLT: performance losses time, minutes
• NOpT: net operating time = OpT-PLT
• pf: performance rate = NOpT/OpT
• QLT: quality losses time, minutes
• VT: valuable time = NOpT-QLT
• qu: quality rate = VT/NOpT
• oee: av $\times$ pf $\times$ qu
• nstops: number of stops
• hum: % of humidity
• temp: temperature in ºC
• wPT: work shift time
• av.level: availability level ([80-100%]: very good, [60-80%): acceptable, [40-60%): improvable, [0-40%): very poor)
• pf.level: performance level
• qu.level: quality level
• oee.level: oee level

MVTF

This folder contains all the code required to run the multivariate and univariate models and reproduce Figures 5.1 and 5.2 of the paper.

• functions: R script with some auxiliary functions that will be used to run the model, including functions for the clustering step (section 4.3 in the paper).
• update_model: R code for parameter estimation and response prediction, including Algorithm 1 and Algorithm 2 of the paper.
• theme_mtf: customized theme for figures.
• run_mv_model: R code to run the multivariate version of the model using a subset of the whole dataset comprising 4 consecutive weeks of clean data. Responses $\mathbf y_n$, covariates $\mathbf x_n$ and classification variables $\mathbf t_n$ are chosen as stated in Section 5. Different model configurations can be tested by the user changing either of them. The performance of the prediction method is measured using a 4-fold cross-validation technique alternatively using one week as the test set, whereas the remaining three weeks are used to train the model.
• run_uv_model: R code to run the univariate version of the multivariate code above.
• mvtf_figures: R code to reproduce Figures 5.1 and 5.2 in the paper.

To ensure the proper working of the code please run the scripts in the following order: run_mv_model - run_uv_model - mvtf_figures.

Computations were run using R-4.2.1.

References

Miguelez, F., Doncel, J. and Ugarte, M.D. (2024). An Adaptive Learning Approach to Multivariate Time Forecasting in Industrial Processes. Submitted. (ArXiv: https://arxiv.org/abs/2403.07554)